AI News — Friday, November 20, 2026
Apple is reportedly working on a proprietary AI agent designed to operate and manage tasks directly on macOS, signaling a deeper integration of AI into its ecosystem.
Anthropic's seven founders collectively sold close to $2 billion in shares during the company's recent IPO, marking a significant financial milestone for the AI safety-focused firm.
Meta's upcoming Muse Charm, a new AI-powered wearable device, is rumored to launch with a price tag of $300, indicating Meta's continued push into consumer AI hardware.
Google has launched Tool Retrieval capabilities within the Gemini API, enabling developers to more effectively integrate external tools and services with their AI agents.
OpenAI details its 'Performance Factory' approach, a systematic method for continuously optimizing and maintaining the speed and efficiency of ChatGPT at scale.
Google has expanded the availability of its Gemini AI app to Android devices with lower memory specifications, increasing accessibility for a broader user base.
This article explores three critical questions regarding AI strategy, drawing parallels to ancient Chinese military tactics to guide modern AI development and deployment.
This piece delves into the practical challenges and necessary steps for successful, widespread AI adoption across an entire team or organization, beyond just individual use cases.
The article critiques the prevalent narratives of AI apocalypse, urging readers to look beyond sensationalism and understand the nuanced realities and underlying details of AI development.
Researchers introduce CurvDelta, a novel modeling approach that uses low-curvature delta representations to enhance learning and 'unlearning' processes in surgical AI applications.
HINT-WAM proposes a new training methodology that leverages hindsight information to improve the performance and robustness of World Action Models, crucial for advanced AI agents.
WeVisForensics presents a self-evolving AI agent designed for comprehensive detection of AI-generated videos across various modalities, addressing the growing challenge of deepfakes.
LightTAR offers a new method for creating lightweight and efficient target-aware representations, significantly improving multimodal tabular learning tasks.
This guide advises on optimizing costs by using simpler decision models for straightforward tasks instead of expensive LLMs, outlining strategies for efficient AI implementation in 2026.
This article explores potential vulnerabilities where seemingly innocuous data, like a plain table, could be exploited to manipulate or 'hijack' the behavior of an AI agent.